CN116861300B - Automatic auxiliary labeling method and device for complex maneuvering data set of military machine type - Google Patents

Abstract

The invention relates to an automatic auxiliary labeling method and device for a complex maneuvering data set of a military machine type. The method comprises the following steps: collecting historical flight parameter sequence data and selecting a standard maneuver sample sequence; preprocessing historical flight parameter sequence and target maneuver template sequence data; using a Matrix Profile data structure, and using a target maneuvering template sequence to preliminarily match in a historical flight parameter sequence and extracting a to-be-identified flight parameter subsequence; pre-classifying the identification subsequence; performing similarity matching on the flight parameter subsequence to be identified and the target maneuver template sequence by adopting a multidimensional dynamic time planning method, and finishing the fine classification of maneuver; and carrying out three-dimensional visualization on the flight parameter subsequence, and manually rechecking the automatically marked maneuver categories to obtain a maneuver data set with the tag. The method can remarkably improve the labeling efficiency and accuracy of the complex maneuvering data set, and is suitable for various military models.

Description

Automatic auxiliary annotation method and device for military aircraft complex maneuver data sets

Technical field

The invention belongs to the technical field of intelligent processing of flight data, and specifically relates to an automatic auxiliary labeling method and device for complex maneuver data sets of military aircraft.

Background technique

Scientifically evaluating the quality of flight training is of vital significance for analyzing pilots' control habits, improving pilots' flight driving skills, and ensuring flight safety. Maneuvering action recognition is a key step in flight training quality assessment, and a large amount of assessment content is based on obtaining specific categories of action sequences. The flight parameter data recorded and saved by the airborne flight parameter system in the form of multi-dimensional time series data provides an objective and scientific basis for maneuver recognition.

At present, maneuver recognition methods based on flight parameter data mainly include methods based on pattern matching, methods based on expert systems, and intelligent methods based on deep learning. Traditional pattern matching-based methods require manual adjustment of thresholds. Expert system-based methods require the establishment of a manual rule knowledge base based on the prior knowledge of domain experts. The accuracy of these two types of methods for identifying complex maneuvers with high similarity needs to be further improved. In recent years, artificial intelligence has developed and progressed rapidly, and methods based on deep learning have shown remarkable performance in the field of motor action recognition. However, deep learning-based methods require large datasets of annotated maneuver samples for model training. In addition, flying parameter data recorded by multi-modal airborne sensors such as gyroscopes or accelerometers contains dozens to hundreds of parameters, which is far more difficult to understand than data from other modal sensors such as cameras. The traditional method of labeling maneuver data sets is to manually observe and analyze the flight data, identify the maneuver segments, and label them. However, this manual labeling method has the following problems: (1) The workload is large, and manual labeling of a large amount of flight data requires a lot of time and human resources; (2) It is highly subjective, and different personnel can identify maneuvers There may be subjective differences in labeling and labeling, resulting in inconsistent results; (3) The accuracy is low. The extraction and category labeling of maneuver sequence segments require professionals. The accuracy of labeling is affected by factors such as individual ability, experience, fatigue, etc. There may be labeling errors. Wrong situation. Therefore, there is a need for a method and device that can automatically assist in labeling maneuvers in flight data to improve labeling efficiency and accuracy.

Contents of the invention

The technical problem to be solved by this invention is to address the deficiencies of the above-mentioned existing technologies and provide a technical solution for automatic auxiliary labeling of complex maneuver data sets for military aircraft to solve the problem of labeling accuracy and efficiency of target maneuver categories under massive flight parameter data. Low technical issues.

In order to achieve the above-mentioned object of the invention, the present invention adopts the following technical solutions:

S1. Collect the historical flight parameter sequence data of pilot flight training, and select the standard maneuver sample sequence as the target maneuver template;

S2. Perform data preprocessing on the collected historical flight parameter sequence data and the selected target maneuver template sequence;

S3. Based on the Matrix Profile data structure, use the target maneuver template sequence to initially match and extract the flight parameter subsequence to be identified in the historical flight parameter sequence that has undergone data preprocessing;

S4. Pre-classify the subsequences to be identified extracted in S3 through obvious features, that is, perform pre-matching of maneuver categories;

S5. Use the MDTW algorithm to perform similarity matching between the flight parameter subsequence to be identified and the target maneuver template sequence in the pre-classified action sequence to complete the subdivision of maneuvers;

S6. Perform three-dimensional visualization of the subdivided flight parameter subsequences, and conduct manual and accurate review of the automatically labeled maneuver category labels to obtain the final labeled maneuver data set.

Preferably, the historical flight parameter sequence data of pilot flight training as described in S1 is collected, and the standard maneuver sample sequence is selected as the target maneuver template; specifically including:

S1-1. Collect the historical flight parameter sequence data of pilot flight training, and decode the historical flight parameter data into table data that can be directly read by the computer;

S1-2. Based on the flight training syllabus and actual needs, determine the type of target maneuver to be marked, and select the flight parameter subsequence that meets the requirements of the maneuver operation specifications in the flight training manual as the standard target maneuver template sequence.

Preferably, S2 performs data preprocessing on the collected historical flight parameter sequence and the selected target maneuver template sequence; specifically including:

S2-1. Get the pitch angle, tilt angle, air pressure height, X-axis angular velocity, Y-axis angular velocity, Z-axis angular velocity, horizontal acceleration and vertical acceleration parameters that can determine different action categories, and compare the historical flight parameter sequence and target maneuver template Missing values in the sequence are filled, and then standardized preprocessing is performed on each dimension of the original data.

The data standardization processing formula is: Among them,/> Represents the mean value of each dimension of data in the flying parameter data,/> Represents the standard deviation of each dimension of data in the flying parameter data. After preprocessing, the average value of each dimension of data is 0 and the standard deviation is 1.

Preferably, based on the Matrix Profile data structure, S3 uses the target maneuver template sequence to initially match and extract the flight parameter subsequence to be identified in the historical flight parameter sequence that has been preprocessed by the data; specifically including: given the historical flight parameter sequence that has been preprocessed by the data. Flight parameter time series , the target maneuver template sequence used for query/> . Use Target Maneuver Template Sequence/> A sliding window of length from a data preprocessed historical flight parameter time series/> The starting position starts to slide, and each time the subsequence within the calculation window is the same as the target maneuver template sequence/> distance, the generated length is/> MatrixProfile, query in Matrix Profile is less than the threshold/> The value of , the position of this value is from/> Match and extract the flight parameter subsequence to be identified/> .

in, Indicates the number of parameters,/> Represents the length of historical flight parameter time series, /> Represents the length of the target maneuver template sequence, usually/> Far smaller than/> .

Preferably, in S4, the subsequences to be identified extracted in S3 are pre-classified through obvious features, that is, pre-matching of maneuver categories is performed; specifically including:

S4-1. The above-mentioned obvious features are height and pitch angle. The pre-classified categories include four categories: quasi-double, quasi-dive and jump, quasi-lift and turn, quasi-sharp turn, quasi-roll, and quasi-circling;

S4-2. The specific pre-matching of the maneuver category is: first, set the threshold and threshold/> ,/> Take the average of the maximum height and minimum height differences in each standard target maneuver template sequence, /> Take the average value of the difference between the maximum pitch angle and the minimum pitch angle in each standard target maneuver template sequence, and then, based on the mean of the absolute value of the first-order difference in height of the subsequence to be identified, set the mean value greater than the threshold/> The sequence is divided into lifting actions, which are less than the threshold/> The sequence is divided into non-lifting actions. In lifting movements, according to the mean value of the first-order difference absolute value of the pitch angle of the sub-sequence to be identified, the mean value is greater than the threshold/> The sequence is divided into bucket-like and dive-like jumps, which are smaller than the threshold/> The sequence is divided into lift and turn types; in non-lift actions, according to the mean of the first-order difference absolute value of the pitch angle of the sub-sequence to be identified, the mean is greater than the threshold/> The sequence is divided into classes of circles, and the mean is smaller than the threshold/> The sequence is divided into quasi-sharp turn and quasi-roll;

Preferably, in the maneuver recognition stage in S5, the MDTW algorithm is used to perform similarity matching between the flight parameter subsequence to be identified and the target maneuver template sequence to complete the subdivision of maneuvers; specifically including:

S5-1. The maneuver recognition stage is to perform similarity matching between the flight parameter sub-sequence to be identified obtained after S4 pre-classification and the existing target maneuver template sequence, and calculate the flight parameter sub-sequence to be identified and C targets respectively. MDTW distance of maneuver template sequence to obtain similarity value sequence ,/> Value is less than threshold , then the action to be recognized is determined as/> Corresponding standard action category. Among them/> Set based on the MDTW distance value between the actual maneuver sequence and the corresponding target maneuver template sequence.

S5-2. The calculation principle of MDTW distance is used to calculate the similarity value of two action sequences. Assume that action sequence 1 is , action sequence 2 is/> ,MDTW path matrix is/> ,/> Represents the dimension of the action sequence, and/> Represents the length of action sequence 1 and action sequence 2 respectively,/> Represents the weight of the corresponding dimension;

The action sequence 1 and action sequence 2 are defined as: Among them,/> It is the fifth step of action sequence 1/> Dimensional features are in Chapter/> The value of a point,/> For the second step of action sequence 2/> Dimensional features are in Chapter/> value of a point.

The MDTW path matrix is defined as: ,/> ,in, For action sequence 1/> Dimensional features are in Chapter/> The value of a point,/> For action sequence 2/> Dimensional features are in Chapter/> value of a point. /> Express/> The weight of the dimension,/> Indicates action sequence 1/> Value vectors of all dimensional features of points and action sequence 2/> The frame-weighted matching distance of the value vectors of all dimensional features of a point;

S5-3, multi-dimensional dynamic time planning distance between action sequence 1 and action sequence 2 Defined as the cumulative distance for the optimal regular path/> The regular path should satisfy the three constraints of boundary, continuity and monotonicity. The formula is defined as:

Among them, the boundary condition is that the starting point of the regular path must be/> , the end point must be/> ; The continuity condition is that the regular path should be continuous, that is, jumping from one point to the next point, which can be to the right, upward, or up-right; the monotonicity condition is that the moving direction of the regular path should be monotonic, that is, not will move in the opposite direction. Construct the optimal regular path according to the above three constraints, and calculate the cumulative distance of the optimal regular path to obtain/> .

Preferably, S6 performs three-dimensional visualization of the subdivided flight parameter subsequences, and reviews the automatically labeled maneuver category labels to obtain the final labeled maneuver data set; specifically including:

S6-1. Use the pitch angle, roll angle and yaw angle information to visualize the flight attitude in a three-dimensional manner using the flight parameter sub-sequences that have been automatically subdivided by the MDTW algorithm. Use altitude, longitude and latitude information to use the altitude, longitude and latitude information to perform the three-dimensional visualization of the flight parameters that have been automatically subdivided by the MDTW algorithm. The sub-sequence performs three-dimensional visualization of the flight trajectory. Based on the three-dimensional visualization results of the flight attitude and trajectory, the automatically labeled maneuver category labels are reviewed. When the classification result of the MDTW algorithm is consistent with the manual classification result, the flight parameter sub-sequence and the corresponding are stored. Label category. When the classification result of the MDTW algorithm is inconsistent with the manual classification result, the manual classification result is used as the category label, and the corresponding flight parameter subsequence is stored, thus obtaining a labeled maneuver data set.

The invention also provides an auxiliary annotation device for military aircraft complex maneuver data sets.

Specifically, an auxiliary annotation device for military aircraft complex maneuver data sets, including:

Data acquisition module, used to obtain flight parameter sequence data;

A preprocessing module, configured to perform data preprocessing on the flight parameter sequence data to generate preprocessed flight parameter sequence data;

The pre-matching module is used to initially match the target action template with the historical flight parameter sequence to obtain the flight parameter sub-sequence to be identified.

The maneuver classification module is used to perform similarity matching calculations on the flight parameter sub-sequence to be identified and the target maneuver template sequence to obtain a detailed classification result of the maneuver.

The maneuvering action three-dimensional visualization module is used to perform three-dimensional visualization of flight postures and trajectories of subdivided maneuvering action sequences, and then perform manual and accurate review of automatically labeled maneuver types.

The technical solution provided by the present invention has at least the following beneficial effects:

1. The present invention utilizes the Matrix Profile data structure and uses maneuver templates to automatically match and extract target maneuver sub-sequences from massive flight parameter data, greatly improving the positioning and sub-sequence fragments of the time period of the target maneuver in massive historical flight parameters. The extraction efficiency reduces the workload of annotators.

2. After pre-classifying the flight parameter sub-sequences to be identified, the present invention uses the MDTW algorithm to calculate the similarity between multi-dimensional and unequal-length maneuver sequences, thereby achieving refined automatic identification of the maneuver types of the flight parameter sub-sequences to be identified. , reducing the calculation amount of MDTW and improving the recognition accuracy and efficiency.

3. The automatic auxiliary labeling method for military aircraft maneuvers proposed by the present invention is a universal method, which can significantly improve the labeling efficiency of complex maneuver data sets, is suitable for a variety of different military aircraft models, and has a wide range of engineering applications. prospect.

In summary, the method and device for automatic auxiliary annotation of maneuver data sets provided by the present invention realize automatic auxiliary annotation of maneuver segments in flight data by combining the Matrix Profile data structure and the MDTW algorithm, and are automated, accurate and efficient. With the advantages of versatility, it can be widely used in flight action analysis and quality assessment in the military aviation field.

Description of the drawings

Figure 1 is an overall flow chart of an automatic auxiliary annotation method for military aircraft complex maneuver data sets according to one embodiment of the present invention.

Figure 2 is a flow chart of maneuver pre-classification processing according to one embodiment of the present invention.

Figure 3 is a schematic diagram of the MDTW regular path according to an embodiment of the present invention.

Figure 4 is a detailed implementation flow chart of automatic auxiliary annotation of maneuver data sets according to one embodiment of the present invention.

Figure 5 is a schematic structural diagram of an automatic auxiliary annotation device for military aircraft complex maneuver data sets according to an embodiment of the present invention.

Detailed ways

Example

In order to make the purpose, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be clearly and completely described below in conjunction with specific embodiments of the present application and corresponding drawings. Obviously, the described embodiments are only some of the embodiments of the present application, but not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of this application.

As shown in Figure 1, an embodiment of the present invention provides an automatic auxiliary labeling method for military aircraft complex maneuver data sets. The method includes:

S1. Collect the historical flight parameter sequence data of pilot flight training, and select the standard maneuver sample sequence as the target maneuver template.

During the flight of a military aircraft, the multi-sensors on the aircraft will monitor a series of parameters related to the flight status. These flight state-related parameters include pitch angle, roll angle, heading angle, angle of attack and sideslip angle, flight altitude, flight speed, angular velocity, acceleration, etc. Therefore, in the process of labeling maneuvers, different maneuvers can be described by in-depth analysis of the change characteristics of the flight parameters of the maneuvers. It should be emphasized that these specific index parameters do not limit the scope of protection of this application. In addition, all maneuver data within the complete flight time period from the start of the military aircraft to the end of the flight can be obtained, and flight parameter sequence data for part of the complete flight time period of the military aircraft can also be selectively obtained. It should be noted that the acquisition time period of flight parameter sequence data does not limit the scope of protection of this application.

Further, the standard flight parameter subsequence is selected as the target maneuver template according to the operating specifications of the target maneuver in the flight training manual. It should be emphasized that in a preferred implementation provided by this application, the acquired historical flight parameter sequence and standard maneuver sample sequence data include pitch angle, roll angle, heading angle, angle of attack and sideslip angle, and flight altitude. , flight speed, angular velocity, acceleration and other flight status related parameters. In addition, the number of target maneuver templates is equal to the number of labeled categories of maneuvers, and this number does not limit the scope of protection of this application.

S2. Perform data preprocessing on the collected historical flight parameter sequence and the selected target maneuver template sequence.

There is redundant information in the flight status parameters included in the acquired original historical flight parameter sequence and the target maneuver template sequence. In order to reduce the data dimension and improve calculation efficiency when making maneuver judgments, some key parameters with action distinction can be selected. .

Furthermore, in a preferred embodiment provided by this application, the pitch angle, tilt angle, air pressure height, X-axis angular velocity, Y-axis angular velocity, Z-axis angular velocity, horizontal acceleration and vertical acceleration that can determine different action categories are taken. Parameters, fill in the missing values in the historical flight parameter sequence and target maneuver template sequence, and then perform standardized preprocessing on each dimension in the original data.

The data standardization processing formula is:

in, Represents the mean value of each dimension of data in the flying parameter data,/> Represents the standard deviation of each dimension of data in the flying parameter data. After preprocessing, the average value of each dimension of data is 0 and the standard deviation is 1.

S3. Based on the Matrix Profile data structure, use the target maneuver template sequence to initially match and extract the flight parameter subsequence to be identified in the historical flight parameter sequence that has undergone data preprocessing.

Given a data preprocessed historical flight parameter time series , the target maneuver template sequence used for query/> . Using Target Maneuver Template Sequences A sliding window of length from a data preprocessed historical flight parameter time series/> The starting position starts to slide, and each time the subsequence within the calculation window is the same as the target maneuver template sequence/> distance, the generated length is/> Matrix Profile, query in Matrix Profile is less than the threshold/> The value of , the position of this value is from/> Match and extract the flight parameter subsequence to be identified/> .

in, Indicates the number of parameters,/> Represents the length of historical flight parameter time series, /> Represents the length of the target maneuver template sequence, usually/> Far smaller than/> ,threshold/> Set manually.

S4. Pre-classify the subsequences to be identified extracted in S3 through obvious features, that is, perform pre-matching of maneuver categories.

The obvious features are height and pitch angle, and the pre-classified categories include four categories: similar to a dolly, similar to a dive and jump, similar to a lift and turn, similar to a sharp turn, similar to a roll, and similar to a hover. It can be understood that the pre-classified categories described here obviously do not constitute a limitation on the specific protection scope of the present application.

Further, as shown in Figure 2, in a preferred implementation provided by this application, the maneuver pre-matching is specifically: first, determine the threshold and threshold/> . Then, according to the mean of the first-order difference absolute value of the height of the subsequence to be identified, the mean is greater than the threshold/> The sequence is divided into lifting and lowering actions, and those that are smaller than the threshold/> Sequences are divided into non-lifting actions. In lifting movements, according to the mean of the first-order difference absolute value of the pitch angle of the sub-sequence to be identified, the mean is greater than the threshold. The sequence is divided into bucket-like and dive-like jumps, which are smaller than the threshold/> The sequence is divided into lifting and turning types; similarly, in non-lifting and descending actions, according to the mean of the absolute value of the first-order difference of the pitch angle of the sub-sequence to be identified, the mean is greater than the threshold/> The sequence is divided into classes of circles, and the mean is smaller than the threshold/> The sequences are divided into quasi-sharp turns and quasi-rolls. Through maneuver pre-matching, maneuver templates that are obviously inconsistent with the subsequence category to be identified are eliminated, thereby reducing the number of matching motion templates, simplifying calculations, and improving work efficiency.

S5. In the maneuver recognition stage, the multi-dimensional dynamic time planning method is used to perform similarity matching between the flight parameter sub-sequence to be identified and the target maneuver template sequence to complete the subdivision of maneuvers.

Specifically, the maneuver recognition stage is to perform similarity matching between the flight parameter subsequence to be identified obtained after S4 pre-classification and the existing target maneuver template sequence, and calculate the flight parameter subsequence to be identified and C target maneuvers by respectively calculating Multidimensional dynamic time planning distance of action template sequence to obtain similarity value sequence , Value is less than threshold/> , then the action to be recognized is determined as/> Corresponding standard action category. Among them/> It is artificially given based on experience.

The calculation of the similarity value of the two action sequences adopts the calculation principle of multi-dimensional dynamic time planning distance. Assume that action sequence 1 is , action sequence 2 is/> ,MDTW path matrix is/> . /> Represents the dimension of the action sequence, /> and/> Represents the length of action sequence 1 and action sequence 2 respectively,/> Represents the weight of the corresponding dimension.

The action sequence 1 and action sequence 2 are defined as:

Among them,/> It is the fifth step of action sequence 1/> Dimensional features are in Chapter/> The value of a point,/> For the second step of action sequence 2/> Dimensional features are in Chapter/> value of a point.

The MDTW path matrix is defined as: ,/> , where,/> For action sequence 1/> Dimensional features are in Chapter/> The value of a point,/> For action sequence 2/> Dimensional features are in Chapter/> The value of a point,/> Express/> The weight of the dimension,/> Indicates action sequence 1/> Value vectors of all dimensional features of points and action sequence 2/> The frame-weighted matching distance of the value vectors of all dimensional features of a point;

S5-3, multi-dimensional dynamic time planning distance between action sequence 1 and action sequence 2 Defined as the cumulative distance of the optimal regular path/> The regular path should satisfy the three constraints of boundary, continuity and monotonicity. The formula is defined as:

Among them, the boundary condition is that the starting point of the regular path must be/> , the end point must be/> ; The continuity condition is that the regular path should be continuous, that is, jumping from one point to the next point, which can be to the right, upward, or up-right; the monotonicity condition is that the moving direction of the regular path should be monotonic, that is, not will move in the opposite direction. Construct the optimal regular path according to the above three constraints, and calculate the cumulative distance of the optimal regular path to obtain/> .

S6. Perform three-dimensional visualization of the subdivided flight parameter subsequences, and conduct manual and accurate review of the automatically labeled maneuver category labels to obtain the final labeled maneuver data set.

Specifically, the flight parameter sub-sequences that have been automatically subdivided by the multi-dimensional dynamic time planning method are used to visualize the flight attitude and trajectory in three dimensions, and the automatically labeled maneuver category labels are manually and accurately reviewed. When the classification result of the multi-dimensional dynamic time planning method is consistent with the manual classification result, the flight parameter subsequence and the corresponding label category are stored. When the classification result of the multi-dimensional dynamic time planning method is inconsistent with the manual classification result, the manual classification result is stored. As a category label, and store the corresponding flight parameter subsequence, thus obtaining a labeled maneuver data set. In summary, in a preferred implementation provided by this application, the detailed implementation flow chart of automatic auxiliary annotation of maneuver data sets is shown in Figure 4.

Please refer to Figure 5, which is an automatic auxiliary labeling device 100 for military aircraft complex maneuver data sets provided by an embodiment of the present application, including:

Data acquisition module 11, used to acquire flight parameter sequence data;

The preprocessing module 12 is used to perform data preprocessing on the flight parameter sequence data to generate preprocessed flight parameter sequence data;

The pre-matching module 13 is used to initially match the target action template with the historical flight parameter sequence to obtain the flight parameter sub-sequence to be identified.

The maneuver classification module 14 is used to perform similarity matching calculations on the flight parameter sub-sequence to be identified and the target maneuver template sequence to obtain a detailed classification result of the maneuver.

The maneuver three-dimensional visualization module 15 is used to perform three-dimensional visualization of flight postures and trajectories of subdivided maneuver sequences, and then perform manual and accurate review of automatically labeled maneuver types.

The present invention draws on the idea of multi-dimensional time subsequence matching query and proposes an automatic auxiliary labeling method for military aircraft complex maneuver data sets, aiming to automatically detect and label target maneuvers from historical flight data. First, the Matrix Profile data structure is used to initially match and extract the target maneuver subsequence using maneuver templates in massive data; secondly, the Multidimensional Dynamic Time Warping (MDTW) method is used to determine the maneuvers of the segmented maneuver subsequences. type; finally, three-dimensional visualization of the flight attitude and trajectory of the maneuver sequence is performed to facilitate manual and accurate review of the automatically labeled maneuver types. This automatic auxiliary annotation method can significantly improve the annotation efficiency of complex maneuver data sets, is suitable for a variety of different military aircraft models, and has broad engineering application prospects.

The above descriptions are only examples of the present application and are not intended to limit the present application. To those skilled in the art, various modifications and variations may be made to this application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of this application shall be included in the scope of the claims of this application.

Claims (4)

1. An automatic auxiliary annotation method for military aircraft complex maneuver data sets, which is characterized in that the method includes: S1. Collect the historical flight parameter sequence data of pilot flight training, and select the standard maneuver sample sequence as the target maneuver template; S2. Perform data preprocessing on the collected historical flight parameter sequence data and the selected target maneuver template sequence; S3. Based on the Matrix Profile data structure, use the target maneuver template sequence to initially match and extract the flight parameter subsequence to be identified in the historical flight parameter sequence that has undergone data preprocessing; S4. Pre-classify the subsequence of flight parameters to be identified extracted in S3 through obvious features, that is, perform pre-matching of maneuver categories; S5. Use the MDTW algorithm to perform similarity matching between the flight parameter subsequence to be identified and the target maneuver template sequence in the pre-classified action sequence to complete the subdivision of maneuvers; S6. Perform three-dimensional visualization of the subdivided flight parameter subsequences, and review the automatically labeled maneuver category labels to obtain the final labeled maneuver data set; As described in step S3, based on the Matrix Profile data structure, the target maneuver template sequence is used to initially match and extract the flight parameter subsequence to be identified in the historical flight parameter sequence that has undergone data preprocessing, specifically including: Given a data preprocessed historical flight parameter time series Target maneuver template sequence used for query/> The sliding window of the length of the target maneuver template sequence Q is used to slide from the starting position of the historical flight parameter time series F after data preprocessing. The distance between the subsequence in the window and the target maneuver template sequence Q is calculated each time, and the generated length is n. Matrix Profile of -m+1, query the value less than the threshold γ in the MatrixProfile. The position of this value is the flight parameter subsequence to be identified that is matched and extracted from F/> Among them, k represents the number of parameters, n represents the length of the historical flight parameter time series, and m represents the length of the target maneuver template sequence. Usually m is much smaller than n; As described in step S6, perform three-dimensional visualization of the subdivided flight parameter subsequences, and perform manual and accurate review of the automatically labeled maneuver category labels to obtain the final labeled maneuver data set, which specifically includes: Use the pitch angle, roll angle and yaw angle information to perform three-dimensional visualization of the flight attitude of the flight parameter sub-sequence automatically subdivided by the MDTW algorithm, and use altitude, longitude and latitude information to fly the flight parameter sub-sequence automatically subdivided by the MDTW algorithm. For three-dimensional visualization of the trajectory, the automatically labeled maneuver category labels are reviewed based on the flight attitude and trajectory three-dimensional visualization results. When the classification results of the MDTW algorithm are consistent with the manual classification results, the flight parameter subsequence and the corresponding label category are stored. When When the classification results of the MDTW algorithm are inconsistent with the manual classification results, the manual classification results are used as category labels and the corresponding flight parameter subsequences are stored, thus obtaining a labeled maneuver data set; As described in step S4, the flight parameter subsequence to be identified extracted in S3 is pre-classified through obvious features, that is, pre-matching of maneuver categories is performed, specifically including: S4-1. The above-mentioned obvious features are height and pitch angle. The pre-classified categories include four categories: quasi-double, quasi-dive and jump, quasi-lift and turn, quasi-sharp turn, quasi-roll, and quasi-circling; S4-2. The specific pre-matching of the maneuver category is as follows: first, set the threshold H and threshold P, H takes the average value of the maximum height and minimum height difference in each standard target maneuver template sequence, and P takes the average value of each standard target maneuver template sequence. The average value of the difference between the maximum pitch angle and the minimum pitch angle in the standard target maneuver template sequence. Then, according to the average value of the first-order difference absolute value of the height of the flight parameter subsequence to be identified, the sequence with the average value greater than the threshold H is classified into the lift category. Actions, sequences smaller than the threshold H are classified into non-lifting actions. In lifting actions, according to the mean value of the first-order difference absolute value of the pitch angle of the flight parameter subsequence to be identified, the sequence with the mean value greater than the threshold P is divided into categories such as bucket-like and bucket-like. Dive and jump, sequences smaller than the threshold P are classified as lift and turn; in non-lift actions, sequences with a mean greater than the threshold P are classified as circling based on the mean of the first-order difference absolute values of the pitch angles of the subsequences to be identified, and the mean is Sequences smaller than the threshold P are classified into sharp turns and rolls. 2. The automatic auxiliary annotation method for military aircraft complex maneuver data sets according to claim 1, characterized in that in step S1, the historical flight parameter sequence data of pilot flight training is collected, and the standard maneuver sample sequence is selected as Target maneuver template, including: S1-1. Collect the historical flight parameter sequence data of pilot flight training, and decode the historical flight parameter data into table data that can be directly read by the computer; S1-2. Based on the flight training syllabus and actual needs, determine the type of target maneuver to be marked, and select the flight parameter subsequence that meets the requirements of the maneuver operation specifications in the flight training manual as the standard target maneuver template sequence. 3. The automatic auxiliary annotation method for military aircraft complex maneuver data sets according to claim 1, characterized in that in step S2, the collected historical flight parameter sequence and the selected target maneuver template sequence are pre-data-prepared. Processing, specifically including: Take the pitch angle, tilt angle, air pressure height, X-axis angular velocity, Y-axis angular velocity, Z-axis angular velocity, horizontal acceleration and vertical acceleration parameters that can determine different action categories, and correct the missing items in the historical flight parameter sequence and target maneuver template sequence. Values are filled, and then normalized preprocessing is performed on each dimension in the original data. 4. The automatic auxiliary annotation method for military aircraft complex maneuver data sets according to claim 1, characterized in that the MDTW algorithm is used in step S5 to perform similarity matching between the flight parameter sub-sequence to be identified and the target maneuver template sequence. , complete the subdivision of maneuvers, including: S5-1. The maneuver recognition stage is to perform similarity matching between the flight parameter subsequence to be identified obtained after S4 pre-classification and the existing target maneuver template sequence. By separately calculating the flight parameter subsequence to be identified and c target maneuvers MDTW distance of the template sequence, the similarity value sequence S = [S ₁ ,..., _Sc ] is obtained. If the S _i value is less than the threshold τ _i ∈τ = [τ ₁ ,..., τ _c ], then the sequence to be The recognized action is determined to be the standard action category corresponding to _Si , where τ is set according to the MDTW distance value between the actual maneuver sequence and the corresponding target maneuver template sequence; S5-2. The calculation principle of MDTW distance is used to calculate the similarity value of two action sequences. It is assumed that action sequence 1 is Action sequence 2 is/> The MDTW path matrix is P, k represents the dimension of the action sequence, m and n represent the lengths of action sequence 1 and action sequence 2 respectively, a[1,...,k] represents the weight of the corresponding dimension; The action sequence 1 and action sequence 2 are defined as: Among them, x _km is the value of the k-th dimension feature of action sequence 1 at the m-th point, and y _kn is the value of the k-th dimension feature of action sequence 2 at the n-th point; The MDTW path matrix is defined as: Among them, x _qi is the value of the q-dimensional feature of action sequence 1 at the i-th point, y _qj is the value of the q-dimensional feature of action sequence 2 at the j-th point, a _q represents the weight of the q dimension, e _ij Represents the frame-weighted matching distance between the value vectors of all dimensional features at the i-th point in action sequence 1 and the value vectors of all dimensional features at the j-th point in action sequence 2; S5-3. The multi-dimensional dynamic time planning distance MDTW(X,Y) of action sequence 1 and action sequence 2 is defined as the sum of the cumulative distance r _ij of the optimal regular path. The regular path should satisfy the three characteristics of boundary, continuity and monotonicity. Constraints, the formula is defined as: MDTW(X,Y)＝min{r _mn } r _ij ＝e _ij +min{r _(i-1)(j-1) ,r _(i-1)j ,r _i(j-1) },0<i≤m,0<j≤n, Among them, the boundary condition is that the starting point of the regular path must be e ₁₁ and the end point must be _emn ; the continuity condition is that the regular path should be continuous, that is, jump from one point to the next point, which can be to the right, Up or to the right; the monotonicity condition is that the movement direction of the regular path should be monotonic, that is, it will not move in the opposite direction. According to the above three constraints, the optimal regular path is constructed, and the cumulative distance of the optimal regular path is calculated to obtain r _ij .